JP2009301125A - Conference voice recording system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a framework that allows participants in a meeting to easily search and correct their own statements using an appropriate authority, while minimizing the operations of the participants. <P>SOLUTION: The speaker of each statement is identified based on the speaker's acoustic feature quantity in speech voice at a meeting, direction information, and the like. A correction authority setting unit 107 provides each statement with an appropriate speech correction authority according to reliability at that time. Also, a voice data search unit 111 and a voice correction unit 112 for easily searching and correcting voice after recording are provided so that a conference participant can correct conference voice after the conference with ease and with an appropriate authority. This makes it possible for the participants to have a free discussion while recording and/or sharing appropriate conference voice with minimum operation. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a conference audio recording system for recording, searching, and sharing the audio in a meeting such as a meeting or brainstorming.

  So far, there has been shown a technique for recording a voice in a conference and creating a minutes from the voice. Japanese Patent Application Laid-Open No. 2004-133867 describes a method of automatically creating a minutes by converting a voice at a meeting into a character by a voice recognition device. Non-Patent Document 1 describes a technique in which a clerk transcribes all statements during a meeting with a dedicated tool and saves it as a minutes.

  Further, in Patent Document 2, in a system in which a plurality of participants edit and share a document, the participant has the right to edit the document for each utterance. It is possible to disseminate free opinions while appropriately managing In this technique, in order to identify the participant's remarks, some terminal for identifying each individual is assumed.

  Patent Document 3 describes a technique for recording a conference using a microphone and granting an approval right to each utterance using speaker identification. If the speaker cannot be identified, the chairman will transcribe it with the consent of everyone. Patent Document 4 describes that recording is interrupted by pressing a specific button when making a statement that is not desired to be disclosed.

JP 2000-112931 A JP 2007-328471 A JP 2000-352995 A Japanese Patent Laying-Open No. 2005-072768 Discussion Mining: Knowledge and Discovery from Minutes, IPSJ 67th National Convention, 2005.

  In Patent Document 1 and Non-Patent Document 1, participant's remarks are treated as official remarks, and there is no framework for managing the authority of participants to correct their remarks or easily correcting their remarks. . In actual meeting situations, it is rather rare that all remarks are treated as official, and such assumptions can hinder participants' free remarks. In particular, in meetings such as meetings and brainstorming, if the purpose is to collect opinions widely and foster knowledge, the original purpose may not be sufficiently achieved.

  Further, in a conference (including a TV conference) in which participants make a meeting with each other, the situation in which each participant holds a dedicated input terminal is inconvenient in the following points. First of all, more than the number of dedicated input terminals cannot participate in the conference. Secondly, a situation where the user can speak only through a dedicated input terminal causes the participants to feel excessive mental stress. Third, the situation of speaking through a dedicated input terminal is very different from the conventional meeting method, and it takes a considerable amount of time for participants to get used to the system. Fourth, it is very expensive to install such a dedicated system. In view of the above, it is considered that a meeting place where each person holds a dedicated input terminal is effective only in a limited environment.

  The method of Patent Document 3 does not require the participant to hold a dedicated input terminal, and the speech is managed, but its purpose is to create an accurate minutes, The point of providing an environment where people can speak freely is not considered. In order to allow participants to speak freely even after recording, it is necessary for each participant to have a function for easily searching and editing their own statements. In addition, this method has a problem in that when the speaker identification fails, the consent of all the members is required uniformly, and the amount of work for all the members until the minutes are prepared is large. Considering daily use of the recording system for meetings, it is desirable that the work after the meeting is minimal, and improvement is also necessary in this respect.

  In the case of the method of interrupting recording by pressing a button, the participant cannot cope with later notice that his / her speech was inappropriate. As a result, there arises a problem that a participant's free speech is hindered.

  As described above, although the conventional technology provides a system for recording and sharing the audio at the meeting, it does not provide a sufficient framework for participants to speak freely. The present invention provides a framework that allows participants to easily search and modify their own statements if necessary, while minimizing the work of meeting participants. In addition, we will provide a framework for making it possible to correct statements with appropriate authority even in meetings where the interests of participants are complex.

  In the present invention, the speaker of each utterance is identified from the speaker acoustic feature amount and direction information of the uttered speech, and an appropriate voice correction right according to the reliability at that time is given to each utterance. Further, by providing a voice search unit and a voice correction unit that can easily search and correct the voice after recording, the conference participants can easily correct the conference voice with an appropriate authority after the conference. This allows participants to discuss freely while recording and sharing appropriate conference audio with minimal work.

  According to the present invention, participants can freely discuss in a system for recording and sharing conference minutes.

Embodiments of the present invention will be described below with reference to the drawings.
FIG. 1 is a functional block diagram showing a configuration example of a conference audio recording / sharing system according to the present invention. The system includes a user management unit 001 for registering conference participants in advance, an audio recording unit 002 that operates when recording conference audio, and conference recording correction / modification that corrects the recording content of the conference after the conference ends. An authentication unit 003 is included.

  Hereinafter, processing in the conference audio recording / sharing system shown in FIG. 1 will be described in order.

  First, the process in the user management unit 001 is shown in the flowchart of FIG. When a user uses this system for the first time, the user registration unit 116 registers information about the user. At this time, a user name and the like are registered. The user's voice is also recorded and stored in the user information holding unit 118. If necessary, register a password or register / issue a user-specific ID card. In addition, the user's face photo and the like can be saved at this time.

  Next, processing in the audio recording unit 002 will be described. A voice recording unit 002 identifies a participant from a participant identification unit 110 that identifies a participant in a conference, a voice input unit 101, a voice recording unit 102 that accumulates input voice, and recorded multi-channel voice. Speaker direction detection unit 103, sound source separation unit 104 that separates input speech for each speaker, speaker acoustic feature extraction unit 105 that extracts a feature amount that expresses speaker characteristics from each separated speech, speaker direction Speaker identification unit 106 that determines the speaker of the speech from the information, speaker acoustic features, and participant information identified by the participant identification unit, and the right to modify or modify the speech based on the reliability of the identified speaker There is a correction right setting unit 107 for setting a transfer certificate, an access right information registration unit 108 for storing the above correction right information, a voice database 109 for storing voice and access right, and a voice indexing unit 119. .

  In this embodiment, the voice input unit 101 accepts synchronous inputs from a plurality of microphones. In addition, the audio recording unit 002 may include an image input unit and an image storage unit. Such an application can be used particularly in a TV conference system.

A flowchart of processing in the voice recording unit 002 is shown in FIG.
Before the conference starts, the participant identification unit 110 first identifies participants of the conference. For this purpose, for example, each participant or the chairman speaks the participant name at the beginning, and the participant is identified by voice recognition. At this time, it is also possible to create a speech recognition dictionary from the user name registered in the user information holding unit 118 in the user registration unit 116. Note that the speech recognition method itself is well known in this technical field, so that the description thereof is omitted.

  In addition, there are a method of issuing an ID card for each speaker and reading the card reader at the time of participation, a method of inputting a conference participant from the keyboard, a method of showing and selecting a participant candidate on a display device, etc. Conceivable. There are also a method for identifying the participant name from the acoustic speaker characteristics of the utterance during the conference and a method for identifying the participant name from the face image.

  Furthermore, when a user who is not registered in the system by the user management unit joins the conference, the user joins the conference with the guest account. It is also conceivable that at this time, the user speaks “I am a guest”, learns the voice quality of the speaker from the voice, and uses it in the speaker identification unit 106 in the subsequent stage.

  Participants in the conference identified above are given the right to view all audio recorded in the conference.

  When the conference actually starts, the system sequentially captures the audio of the conference from the audio input unit 101 having a plurality of microphones and stores it in the audio database 109. Each microphone is installed in a predetermined arrangement such as a straight line or a circumference, and the input from each microphone is synchronously passed to the speaker direction detecting unit 103 through a dedicated A / D board. The speaker direction detection unit 103 detects the direction of the sound source from the multi-channel sound. In this case, a plurality of speakers may speak at the same time, and it is desirable that all sound source directions can be accurately detected even in such a case.

Here, the number of microphone elements is M = 2, and a signal obtained from each microphone is represented as x _i (τ) (i = 1, 2). First, short-time Fourier transform is performed on each x _i (τ), and the result is defined as X _i (f, τ). Here, f represents a frequency, and τ represents a frame index of a short-time Fourier transform.

  For each time frequency obtained, the phase difference θ with respect to the 0th microphone input is estimated.

The sound source direction γ is derived by the following equation.

  Here, r represents the distance between the microphone 1 and the microphone 2, and v represents the speed of sound. The direction of the sound source is indicated by an angle when the bisected linear direction of the microphones 1 and 2 is 0 (degree).

  After obtaining the above for each time frequency band, a plurality of sound sources can be localized by creating a histogram with the sound source direction on the horizontal axis and performing a peak search.

  The example in the case where the number of microphone elements M = 2 is shown above, but even when the number of microphones is larger than 2, it can be dealt with by extending the algorithm. Further, methods such as “Makoto Togami et al .: Localization performance evaluation of sound source localization method SPIRE based on sequential phase difference correction processing, 2007 Spring Acoustic Society, 2007” with improved localization accuracy can be used. These details are well known to those skilled in the art and will not be described here.

  Further, the sound source separation unit 104 separates the sound for each sound source based on the direction information obtained above. This can be realized, for example, by using a minimum dispersion beamformer. Of course, other sound source separation methods such as independent component analysis can be used instead.

In the minimum dispersion beamformer, X _i (f, τ) is multiplied by a linear filter w (f) obtained by the following expression to emphasize the sound in the target direction and suppress other sounds.

Here, a (f) represents a spatial transfer characteristic in the target sound direction, and R (f) represents a spatial correlation matrix.
Further details of this process are well known to those skilled in the art and will not be described here.
By the sound source separation process described above, it is possible to obtain a sound separated for each utterance even in a situation where a plurality of people speak at the same time. The following processing is performed separately for each voice segment, with each separated voice as a voice segment.

  First, the speaker acoustic feature extraction unit 105 extracts a speaker acoustic feature amount exhibiting speaker characteristics for each separated speech segment X. Here, MFCC (Mel Frequency Cestrum Coefficient) or the like can be used as the speaker acoustic feature amount. The details of this feature amount are well known to those skilled in the art, and thus the description thereof is omitted.

Next, the speaker identification unit 106 determines the speaker based on the speaker acoustic feature amount, the speaker direction obtained by the speaker direction detection unit, and the participant information obtained from the participant identification unit. In this embodiment, speaker identification using GMM (Gaussian Mixture Model) is used. When a sequence of speaker acoustic features X = {X ₁ ,..., X _n } is given, the likelihood that it is speaker A is expressed as follows.

Here, m _j , v _j , and λ _j are the average of the first normal distribution, the mixture weight of the variance and the distribution, respectively, and the values are learned from the voice of the speaker A held in the user information holding unit in advance. It was.

In order to obtain the acoustic reliability CM _AC (A | X) that the voice is the speaker A, the likelihood P (X | GMM) of the GMM that expresses general acoustic information, called a back-ground model. _bg ) and calculate the ratio to the speaker likelihood.

In addition to the above, the reliability CM _DOA (A | X) that the speech is the speech of the speaker A can be calculated based on the speaker direction estimated by the speech direction detection unit. For this purpose, for example, information that the speaker A often sits in a specific seat is expressed as a probability P (D | A), and it can be obtained as follows.

Here, D represents the voice arrival direction.
As a method of obtaining P (A | D), it is conceivable to obtain the following based on the speech utterance set obtained in the previous conference or the conference currently being recorded.

The above sum of X is the sum for the set of audio segments obtained in the previous conference or the conference currently being recorded. Further, a is an individual user who is an element of A.
From the CM _AC (A | X) and CM _DOA (A | X) obtained above, the reliability CM (A | X) that the voice is the utterance of the speaker A is obtained. For example, with the linear sum above

と表現することができる。この信頼度の値が大きいほど、当該音声セグメントが当該話者の発言である確率が高いと判断できる。

It can be expressed as It can be determined that the greater the reliability value, the higher the probability that the speech segment is the speaker's speech.

In the above, only two of the reliability CM _AC (A | X) obtained from the speaker acoustic information and the reliability CM _DOA (A | X) obtained from the speaker direction information are used as an example. The system includes an imaging unit such as a camera for imaging a conference hall or a conference participant, and based on the image obtained from the imaging unit, the feature information of the speaker's face image is used for each speaker of the utterance. It is also possible to obtain a combination of reliability. Similarly, the reliability for each speaker of the utterance can be obtained from the face image of the speaker and the direction information.

The correction right setting unit 107 assigns a correction right for each speaker to each voice based on the speaker reliability obtained as described above. Here, different correction rights are assigned in the following (1) to (3).
(1) the speaker reliability predetermined threshold theta ₁ or more speakers to impart modifications right to the speaker if there is only one person (2) speakers reliability threshold theta ₁ or more speakers If there is more than one, give a certificate of assignment of correction rights to all the speakers.
(3) When there is no speaker whose speaker reliability is greater than or equal to the threshold value θ _{1 A} certificate of assignment of the right to revision is assigned to all participants.

  Here, the correction right is the right to correct or delete the contents of the audio segment X. The correction right assignment certificate is a right to obtain a correction right when the correction right assignment certificate of all users having the correction right assignment certificate is received.

  If the speaker determined as (1) is a guest account, a certificate of assignment of correction rights is given to all users participating in the conference. Alternatively, it is possible to predetermine a chairperson and give a correction right to the user.

  The flow of the above processing is shown in FIG. When an audio segment is input, it is estimated which conference participant the speech is, and the reliability is calculated. Thereafter, according to the rules (1) to (3), a correction right or a correction right assignment certificate is issued.

  Finally, the access right information registration unit 108 stores the above-mentioned correction right / correction right assignment certificate and the viewing right given to the conference participant in the audio database 109. Here, the direction information obtained by the speaker direction detection unit 103 and the speaker information obtained by the speaker identification unit 106 can also be stored together.

  An example in which these pieces of information are stored in the voice database 109 is shown in FIG. Here, the audio file ID is an identifier uniquely assigned for each recording of the conference, and the audio segment ID is an identifier uniquely assigned for each audio segment. In FIG. 5, the speaker obtained by the speaker identification unit 106 and its reliability are stored in the column “speaker”. In the column “direction”, the speaker direction obtained from the speaker direction detection unit 103 is stored. Such information can be used in the audio data search unit 111 described later.

  In addition to the above, the audio file ID and conference name, participants, recording date and time, and data as shown in FIG. 6 showing the storage location of the file on the storage device are also stored.

  Next, in the audio indexing unit 119, for each audio segment X, the audio data search unit 111 creates a database for searching the audio data by keywords. There are various known techniques for searching a speech database. Here, a method using large vocabulary continuous speech recognition will be described.

  First, the speech indexing unit 119 converts the speaker-identified speech segment X into a word string using a large vocabulary continuous speech recognizer. Each word in the word string is given the reliability output from the large vocabulary continuous speech recognizer. In addition, since the technique of large vocabulary continuous speech recognition is well-known to those skilled in this field, description is abbreviate | omitted.

  Next, index data expressing in which audio file ID / audio segment ID a certain word appears is created from the obtained word string. An example of this is shown in FIG. Here, a triplet of {voice file ID / voice segment ID / reliability output from voice recognition} corresponding to a certain word is created as an index. For example, it can be seen that the word “product” is uttered with a reliability of 0.8 in the audio file ID 0012 and the audio segment ID 0003, and is uttered with a reliability of 0.5 in the audio file ID 0010 and the audio segment ID 0001.

  As a result, the user can use the keyword in the voice data search unit 111 (to be described later) to obtain the voice file in which the keyword is uttered and its segment position.

  The voice indexing unit stores the index data created above in a voice database.

  In addition to the above, as a method of searching a speech database using keywords, a method of creating a word lattice using a large vocabulary continuous speech recognizer, or a subword speech recognizer in units of subwords smaller than words is used. There are known search methods, etc., and these can be used instead. Further, when searching for a keyword, it is possible to deal with processing when a plurality of keywords are input or a compound word is input. Since this technique is known to those skilled in the art, a description thereof will be omitted. The above is the processing in the audio recording unit 002.

  Next, processing in the recording correction / authentication unit 003 that searches, edits, and publishes the voice recorded by the conference participants will be described.

  The recording correction / authentication unit 003 includes a user authentication unit 117 that performs user authentication, a voice data search unit 111 that can search voice data from a keyword, a speaker, and the like, and corrects only voice data for which the user has a correction right. A voice correction unit 112 that can be deleted, and a voice public authentication unit 113 that can perform public authentication only for voices for which the user has public rights. Furthermore, a user with a correction right assignment certificate has a correction right assignment request unit 114 for requesting assignment of a correction right, and a correction right assignment approval unit 115 for approving the correction right assignment request.

  First, the user inputs user-specific information in the user authentication unit 117 in order to view, modify, and approve the conference recording. The user authentication unit 117 identifies the user who is operating the system from the input information and the information stored in the user information holding unit 118. Here, it is possible to allow the user to input a password, and it is also possible to use a more advanced authentication technique such as finger vein authentication. It is also possible to issue an ID card for each user in the user registration unit 116 and use it for authentication.

  When a user accesses the system through the user authentication unit, the user selects viewing approval / modification / revision right transfer request / revision right transfer approval for the conference in which he / she participated, based on the interface shown in FIG. can do.

  A user identified by the participant identification unit 110 in a certain meeting is granted viewing rights to all the voices in the meeting, and thus can view and search the voices. In this case, when “view conference from list” 202 in FIG. 8 is clicked, a list of conferences that can be viewed by itself is displayed as shown in FIG. 9, and the content can be heard. At this time, the conference to which the audio viewing right during the conference is not granted is not displayed and cannot be viewed. A user who has participated in the conference can freely name the conference audio. This name can be set individually for each user, or can be automatically shared among conference participants.

  In the conference voice search 203 of FIG. 8, the conference contents can be searched using the conference name and keywords. A keyword to be searched is input to the text box 208 in FIG. 8 and the search button 209 is pressed. If only one of the meeting name or keyword search is desired, the check box 210 of the item that is not desired to be used may be cleared. When the search button is pressed, the voice data search unit 111 operates to display a list of corresponding files to the user.

  Here, the search by the conference name and the keyword is shown, but it is also possible to perform a search by the speaker or a search based on the direction of the speaker. Further, when images during the meeting are accumulated, a search based on the face image stored in the user information holding unit 118 is also possible.

  If a user who has participated in the conference determines that the remark that the correction right and the right to transfer the correction have been given to the user may be disclosed, the voice public approval unit 113 approves the public release of the user's utterance. Do. At this time, for example, an interface as shown in FIG. 10 may be provided, and public approval may be performed for individual utterances, or all utterances may be publicly approved collectively. If you want to approve the release of individual audio, check only the audio segment you want to make public, and then press the “Approve approved audio for release” button. If you want to publicly approve all utterances, check “Check all voices” first, and all voice segments will be checked.

  An utterance in which a modification right is given to a certain user can be viewed and searched by a user who has not participated in the conference when the user approves the disclosure. An utterance in which a correction right assignment certificate is given to a plurality of users can be viewed and searched by users who have not participated in the conference when all the users holding the correction right assignment certificate have made public approval.

  A user who has participated in the conference and has determined that there is an utterance that needs to be edited when publishing, clicks on conference audio correction 205 in FIG. 8 to activate the audio correction unit 112. The voice correction unit 112 includes a search interface similar to that shown in FIG. 8, and can also search for voice by keyword or search by conference name.

  The behavior of the user after finding the voice segment to be edited varies depending on which access right of the voice segment the user holds. If the user holds the right to correct the voice segment, an operation such as deleting the voice or masking unnecessary parts is performed.

  Here, the voice correction unit 112 has an interface as shown in FIG. 11, for example, and designates the start point and end point of the section to be corrected by dragging the mouse. Further, it is possible to cut out only the keyword part in the voice by inputting the keyword. When a keyword is entered in the text box 301, the start point and end point of the keyword section are set. Since this technique is called word spotting and is well known to those skilled in the art, details are not described here.

  Clicking “Masking specified section” for the section specified above will replace the specified section with white noise or beep sound. If the specified section is deleted, the specified section is deleted.

  The correction result is reflected when the user views the conference voice, but the actual voice database itself can be prevented from being corrected. In this case, it is also possible to perform an operation such as restoring the voice waveform by the system administrator authority.

  The above is the processing when the user holds the right to correct the audio segment. If the user does not have the right to edit the voice that he / she wants to edit, the voice cannot be corrected as it is. In this case, the user clicks the correction right transfer request 206 in FIG. 8 to activate the correction right transfer request unit 114.

  The correction right transfer request unit 114 has an interface as shown in FIG. 12, and notifies all the users holding the sound correction right transfer certificate of the audio segment. At this time, for example, the mail system and this system may cooperate, and a participant who has been notified of the request to transfer the correction right may be notified by email. In addition, the message of user A may be added to the request for assignment of correction rights.

  The user who has received the correction right assignment request notification accesses the system from the user approval unit 117, and then clicks the correction right assignment approval 207 in FIG. 8 to activate the correction right assignment approval unit 115. The modification right transfer approval unit 115 has an interface as shown in FIG. 13, and can confirm the request user name, the requested voice and the message from the requester. If necessary, it is preferable to provide an interface as shown in FIG. 13 for viewing a specified section so that the context before and after the voice can be confirmed.

  If you listen to the audio and decide that you may give the requesting user the right to correct the audio, click the “Approve Approval for Audio Correcting Rights” button to issue the audio transfer right certificate to the requesting user. To do.

  When all the users having the correction right assignment certificate have issued the correction right assignment certificate to the user A, the correction right assignment approval unit 115 gives the user A the right to correct the sound. Thereby, the user A can correct / delete the sound.

  The above is a framework for viewing, searching, correcting, and publishing conference audio. In this framework, only a few conference participants who are not sure whether or not they uttered the voice need to respond to the request for issuance of the right to transfer correction from User A, and the majority of other users are involved in the process. This eliminates the need for the user and greatly reduces the user's labor as a whole. In addition, since the right to modify and access rights are managed for each utterance, it is possible to avoid concerns about improperly modifying each other's voice after users with different interests speak, and to discuss more freely. It becomes possible.

  In the above example, when the user with the right to modify approves the release of the utterance, other users can listen to the sound. It is also possible to publish the voice when the user has approved the release of all the voices.

  FIG. 14 shows the hardware configuration of the above system. The system includes a computer including a CPU and a memory, and the computer includes a voice input device, a data storage device, a keyboard, and a display device. All of the functional units 101 to 119 shown in FIG. 1 are stored in the memory of the computer. When accepting image input, the image input device is also connected to the computer.

  Further, FIG. 15 shows a hardware configuration when this system is combined with a TV conference system. Here, the voice input device and the image input device are dispersed in a plurality of bases, and are greatly different from FIG. 14 in that they are connected to a computer via a network.

The functional block diagram which shows the structural example of the system by this invention. The flowchart which shows the process in a user management part. The flowchart which shows the process in an audio | voice recording part. The flowchart which shows the process sequence of access right setting. The figure which shows the example of a storage of an audio | voice file and audio | voice segment information. The figure which shows the example of storage of audio | voice file information. The figure which shows the example of the index data produced by the audio | voice indexing part. The figure which shows the example of a user screen after user authentication. The figure which shows the list display example of a meeting audio | voice. The figure which shows the interface of public approval. The figure which shows the interface of a voice correction part. The figure which shows the interface of the correction right transfer request part. The figure which shows the interface of the correction transfer assignment part. The figure which shows the hardware structural example of a system. The figure which shows the hardware structural example at the time of combining with a video conference system.

Explanation of symbols

001: User management unit 002: Voice recording unit 003: Recording correction / authentication unit

Claims (9)

An audio recording unit for recording audio;
A speaker identification unit for identifying a speaker from the input voice;
A correction right setting unit that gives a participant a different type of correction right for each utterance of the input voice according to the reliability of speaker identification by the speaker identification unit;
A user authentication unit for authenticating a user;
A conference voice recording system, comprising: a voice correction unit that allows a user who has been given a correction right by the correction right setting unit to correct an utterance given the correction right.   2. The conference voice recording system according to claim 1, wherein the correction right setting unit determines that the reliability of the result of speaker identification as a speaker is higher than a predetermined threshold value when there is only one speaker. A conference characterized by granting the right to correct the utterance and issuing a certificate of transfer of the right to correct the utterance to the plurality of speakers when there are a plurality of speakers whose speaker identification reliability is higher than the threshold value Voice recording system.   In the conference voice recording system according to claim 1, the correction right setting unit, when there is no speaker whose reliability as a result of speaker identification of a certain utterance is higher than a predetermined threshold, A conference audio recording system characterized by issuing a certificate of assignment of correction rights to a speaker.   4. The conference voice recording system according to claim 2 or 3, wherein the right of correcting the utterance is given to a user who has been issued a transfer right transfer certificate from all users having the transfer right transfer certificate. Recording system.   The conference voice recording system according to claim 1, further comprising a voice search unit that allows a participant to search for a voice by a keyword or a speaker name.   2. The conference voice recording system according to claim 1, further comprising a participant identifying unit for identifying a conference participant.   The conference voice recording system according to claim 6, wherein the participant identification unit identifies a participant based on a result of recognizing a voice spoken during the conference.   The conference audio recording system according to claim 1, comprising either or both of an imaging unit that images a conference hall and a speaker direction detection unit, wherein the speaker identification unit is output from the speaker direction detection unit. Speaker identification is performed from either or a combination of an acoustic feature amount representing a speaker direction detection result and speaker characteristics and an image feature amount representing speaker characteristics obtained from an image captured by the imaging unit. Meeting voice recording system.   9. The conference voice recording system according to claim 8, wherein a plurality of microphones are used as voice input units.

Images